Myung - Cognitive Modeling and Robust Decision Making - Spring Review 2012
1. Cognitive Modeling and
Robust Decision Making
5 March 2012
Jay Myung
Program Manager
AFOSR/RSL
Integrity Service Excellence Air Force Research Laboratory
13 July 2012 DISTRIBUTION A: Approved for public release; distribution is unlimited. 1
2. 2012 AFOSR SPRING REVIEW
NAME: Jay Myung, PhD Years as AFOSR PM: 0.75
BRIEF DESCRIPTION OF PORTFOLIO:
Support experimental and formal modeling work in:
1) Understanding cognitive processes underlying human performance in
complex problem solving and decision making tasks;
2) Achieving maximally effective symbiosis between humans and machine
systems in decision making;
3) Creating robust intelligent autonomous systems that achieve high
performance and adapt in complex and dynamic environments.
LIST SUB-AREAS IN PORTFOLIO:
• Mathematical Modeling of Cognition and Decision (3003B)
• Human-System Interface and Robust Decision Making (3003H)
• Robust Computational Intelligence (3003D)
DISTRIBUTION A: Approved for public release; distribution is unlimited. 2
3. Program Roadmap
Adaptive, Natural or Artificial, Intelligence as Computational
Algorithms Requiring Interdisciplinary Approach
General purpose algorithms that the brain uses to
achieve adaptive intelligent computation.
Behavior/cognition
M
I
Computation N
D Mind as adaptive computational
Neuroscience algorithm (software) running
on the brain (hardware)
Behavior/Cognition: extracting and defining the problem
to be solved, consolidated as robust empirical laws.
Computation: solves a well-defined problem, in terms
of optimization, estimation, statistical inference, etc.
Neuroscience: implementing the solution by the neural
architecture (including hardware and currency).
DISTRIBUTION A: Approved for public release; distribution is unlimited. 3
4. Computational Algorithms for
Cognition and Decision
Challenge for an Adaptive Intelligent System to Solve:
Massive Useful Actionable Robust
Raw Data Information Knowledge Decision
(1) Neuronal signal processing algorithms: optimal balance
between representation (encoding) and computation (decoding)
in information processing of the brain.
(2) Causal reasoning and Bayesian algorithms: optimal fusion
of prior knowledge with data/evidence for reasoning and
prediction under uncertainty, ambiguity, and risk.
(3) Categorization/classification algorithms: optimal generalization
from past observations to future encounters by regulating the
complexity of classifiers while achieving data-fitting performance.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 4
5. Program Trends
• Memory, Categorization, and Reasoning
• Optimal Planning/Control and Reinforcement Learning
• Mathematical Foundation of Decision Under Uncertainty
• Causal Reasoning and Bayesian/Machine Learning Algorithms
• Neural Basis of Cognition and Decision
• Computational Cognitive Neuroscience
• Computational Principles of Intelligence and Autonomy
• Cognitive Architectures
• Software System Architectures and Sensor Networks
DISTRIBUTION A: Approved for public release; distribution is unlimited. 5
6. Mathematical Modeling of Cognition and
Decision (3003B)
Goal:
Advance mathematical models of human performance in
attention, memory, categorization, reasoning, planning, and decision
making.
Challenges and Strategy:
• Seek algorithms for adaptive intelligence inspired by brain science
• Multidisciplinary efforts cutting across mathematics, cognitive
science, neuroscience, and computer science.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 6
7. Neuronal Basis of Cognition
(Aurel Lazar, Columbia EECS)
Neuronal Signal Processing
(Hodgkin & Huxley, 1963, Nobel Prize)
Prof. Aurel
Lazar
“Cognition is an End-product of Neural Computation.”
DISTRIBUTION A: Approved for public release; distribution is unlimited. 7
8. Neuronal Basis of Cognition
(Aurel Lazar, Columbia EECS)
The Scientific Challenge: The Holy Grail of Neuroscience
- What is the neural code?
- Can we “reconstruct” cognition from neural spiking data?
?
DISTRIBUTION A: Approved for public release; distribution is unlimited. 8
9. Neuronal Basis of Cognition
(Aurel Lazar, Columbia EECS)
Neural Population Coding Hypothesis “Brain as A/D and D/A Signal Converters”
DISTRIBUTION A: Approved for public release; distribution is unlimited. 9
10. Neuronal Basis of Cognition
(Aurel Lazar, Columbia EECS)
Video Stream Recovery from Spiking Neuron Model
• First-ever demonstration of natural scene recovery from spiking neuron models
based upon an architecture that includes visual receptive fields and neural
circuits with feedback.
• Scalable encoding/decoding algorithms were demonstrated on a parallel
computing platform.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 10
11. Human System Interface and Robust Decision
Making (3003H)
Goal:
Advance research on mixed human-machine systems to aid
inference, communication, prediction, planning, scheduling, and
decision making.
Challenges and Strategy:
• Seek computational principles for symbiosis of mixed human-
machine systems with allocating and coordinating requirements.
• Statistical and machine learning methods for robust reasoning and
strategic planning.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 11
12. Flexible, Fast, and Rational Inductive
Inference (Griffiths YIP, Berkeley Psy)
The Problem:
• Understand how people are capable of fast, flexible and rational
inductive inference.
Prof. Thomas
Griffiths
The Scientific Challenge:
Develop a framework for making automated systems capable of solving inductive
problems (such as learning causal relationships and identifying features of images)
with the same ease and efficiency of humans.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 12
13. Flexible, Fast, and Rational Inductive
Inference (Griffiths YIP, Berkeley Psy)
A Bayesian Statistical Approach:
Analyze human inductive inference from the perspective of Bayesian statistics.
Explore Monte Carlo algorithms and nonparametric Bayesian models as an
account of human cognition. Test models through behavioral experiments.
Features:
Features are the elementary primitives of
cognition, but are often ambiguous.
Inferring a feature representation is an
inductive inference problem.
Challenge: How do you form a set of
possible representations?
DISTRIBUTION A: Approved for public release; distribution is unlimited. 13
14. Flexible, Fast, and Rational Inductive
Inference (Griffiths YIP, Berkeley Psy)
Nonparametric Bayesian Algorithms: Multi-level Category Learning
Basic idea: Use flexible hypothesis spaces
from nonparametric Bayesian models with
potentially infinite many features.
Learning by examples: Combines structure
of a bias towards simpler feature
representations, but with the flexibility to grow
in complexity as more data is observed.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 14
15. Robust Computational Intelligence (3003D)
Goal:
Advance research on machine intelligence architectures that derive
from cognitive and biological models of human intelligence.
Challenges and Strategy:
• Seek fundamental principles and methodologies for building
autonomous systems that learn and function at the level of flexibility
comparable to that of humans and animals.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 15
16. Practical Reasoning in
Robotic Agents (Pagnucco, UNSW CS)
Objective: Extract summarized, actionable
knowledge from raw data in complex problem
domains (e.g., spatial mapping).
DoD benefits: Enables autonomous systems
to automatically adapt to unfamiliar situations
in the presence of erroneous information.
Technical approach: Formal A.I. methods
for integrating logical and non-logical
reasoning in robotic agents (e.g., MAVs).
Spatial Mapping
DISTRIBUTION A: Approved for public release; distribution is unlimited. 16
17. Modeling Functional Architectures of
Human Decision Making (Patterson, AFRL/RH)
Objective: Understand functional Intuitive Decision Making:
architecture of human decision making. Synthetic Terrain Imagery & Object Sequences
DoD benefits: Increased knowledge about
designs of human-machine systems.
Technical approach: Combine analytical
and intuitive decision making within formal
tests of functional architecture.
Hummer
Truck
Rocket
Truck S1 S2 Launcher
IN OUT
S0 Tank Truck S5
Patriot Rocket
Launcher S3 S4 Launcher
Patriot
Launcher
Tank
DISTRIBUTION A: Approved for public release; distribution is unlimited. 17
18. Joint Initiative with AFRL/RH in
Neuroergonomics Research
(Parasuraman, George Mason U Psy)
CENTEC: Center of Excellence in Neuro-Ergonomics,
Technology and Cognition (since July 2010)
Goal: Support the Air Force mission of enhanced human
Prof. Raja effectiveness through
Parasuraman • Research in neuroergonomics, technology, and cognition
• Training of graduate students and postdoctoral fellows
• Collaborations with scientists of AFRL/RH
• “Direct transitions” of university research to AFRL
NEUROADAPTIVE LEARNING, NEUROIMAGE, BEHAVIORAL GENETICS, ATTENTION,
TRUST IN AUTOMATION, MULTI-TASKING, MEMORY, SPATIAL COGNITION
DISTRIBUTION A: Approved for public release; distribution is unlimited. 18
19. Other Organizations that Fund
Related Work
• ONR (PM: Paul Bello)
• Representing and Reasoning about Uncertainty Program
• Theoretical Foundations for Socio-Cognitive Architectures
Program
• ONR (PM: Tom McKenna)
• Human Robot Interaction Program
• ONR (PM: Marc Steinberg)
• Science of Autonomy Program
ARO (PM: Janet Spoonamore)
• Decision and Neurosciences Program
NSF (PMs: Betty Tuller & Lawrence Gottlob)
• Perception, Action & Cognition Program
• NSF (PM: Sven Koenig)
• Robust Intelligence Program
DISTRIBUTION A: Approved for public release; distribution is unlimited. 19
20. Recent Highlights
CENTEC team (PI: R. Parasuraman, GMU)
- Special issue in Neuroimage on
neuroergonomics
- Joint papers by GMU and AFRL/RH
scientists
Prof. T. Walsh team (NICTA, Australia)
- Eureka Prize for Peter Stuckey
(Highest prize in CS in Australia)
- Best Paper Prize @ AAAI 2011
DISTRIBUTION A: Approved for public release; distribution is unlimited. 20
21. Program Summary
Transformational Impacts and Opportunities
The Basic Premise of Investment Philosophy
Cognition/Intelligence as computation algorithm:
Requires multidisciplinary research efforts to uncover brain
algorithms, from computer science, mathematics, statistics,
engineering, and psychology.
1. Neurocomputational foundation of cognition
• We stand “almost” at the threshold of a major scientific
breakthrough reminiscent of genetics in the 1950s, i.e.,
Watson & Click (1953; Nobel Prize, 1962).
2. Self-learning decision support systems
• Machine learning algorithms for inference and decision
making capability comparable to that of humans.
3. Trust-worthy autonomous agents
• Capable of sense-making massive raw data for
prediction, planning, communication, and decision.
DISTRIBUTION A: Approved for public release; distribution is unlimited. 21
22. Questions?
Thank you for your attention
Jay Myung, Program Manager, AFOSR/RSL
Jay.myung@afosr.af.mil
DISTRIBUTION A: Approved for public release; distribution is unlimited. 22
23. Back-up Slides to Follow
DISTRIBUTION A: Approved for public release; distribution is unlimited. 23
24. Neuronal Basis of Cognition
(Aurel Lazar, Columbia EECS)
Six Disruptive Basic Research Areas
(Honorable Lemnios, Asst Sec Def. (R&E))
1. Metamaterials and 3. Cognitive Neuroscience:
Plasmonics
More deeply understand and more fully
2. Quantum Information exploit the fundamental mechanisms of
Science
the brain.
3. Cognitive Neuroscience
4. Nanoscience and • Key Research Challenges:
Nanoengineering - Solving the inverse problem of
5. Synthetic Biology predicting human behavior from
6. Computational Modeling brain signals
of Human and Social - Translating clinical measurements &
Behavior
analyses to uninjured personnel
- Developing models incorporating
individual brain variability
DISTRIBUTION A: Approved for public release; distribution is unlimited. 24
25. IMPLICIT LEARNING of ARTIFICIAL GRAMMAR
(e.g., Reber, 1967)
Rules for „sentence‟ construction: Finite State Algorithm
P
T
T S1 S2 S
IN OUT
S0 S5
X P
V S3 S4 S
V
X
Sentences of length 6-8 used (34 total possible); examples: TPTXVPS; VXXVPS
Participants learned the grammar with fewer errors than participants who learned
random letter strings—but few could provide answers about the rules
Pattern 'chunks' can be abstracted via a rudimentary inductive process
without explicit strategies
Distribution A
DISTRIBUTION A: Approved for public release; distribution is unlimited. 25
26. CHARACTERISTICS OF 2 PROCESSES
(derived from Evans, 2008)
“System 1” (Intuitive) System 2” (Analytic)
*Situational Pattern Recog *Deliberative
Unconscious Conscious
Implicit Explicit
Automatic Controlled
Low effort High effort
Rapid Slow
High capacity Low capacity
Holistic, perceptual Analytic, reflective
Domain specific (inflexible) Domain specific & general (flexible)
Contextualized Abstract
Nonverbal Linked to language
Independent of working Limited by working memory capacity
memory
Distribution A
DISTRIBUTION A: Approved for public release; distribution is unlimited. 26